OFDM-Based Neuromorphic Wake-Up Signal

arXiv:2610.36914v1 [cs.NI] 27 Oct 2026

OFDM-Based Neuromorphic Wake-Up Signal Gaspard Goupy

Elias Raninen

Dani Korpi

Nokia Bell Labs Espoo, Finland gaspard.goupy@nokia-bell-labs.com

Nokia Bell Labs Espoo, Finland elias.raninen@nokia-bell-labs.com

Nokia Bell Labs Espoo, Finland dani.korpi@nokia-bell-labs.com

Abstract—Wake-up radios (WURs) reduce energy consumption by continuously monitoring the channel for a wake-up signal (WUS) while the main receiver remains asleep. Neuromorphic computing offers a promising approach to ultra-low-power WUS detection through sparse, event-driven processing. However, realizing these benefits requires spike-like inputs, which conventional orthogonal frequency-division multiplexing (OFDM) waveforms do not naturally provide. To enable neuromorphic WUS detection in OFDM-based systems, we introduce a neuromorphic wakeup signal (nWUS) that can be generated by a standard OFDM transmitter. The proposed design superposes frequency-domain vectors with linear phase progressions across the OFDM subcarriers, producing a time-domain waveform with sparse amplitude peaks at predefined positions. These peaks encode the wake-up signature in a spike-like representation that facilitates neuromorphic processing at the receiver. Building on this structure, we then propose a noncoherent neuromorphic detector that encodes received peaks into spikes and performs sparse correlation only upon spike events. Simulations over a 3GPP urban macro channel show that the proposed approach outperforms on-off keying (OOK) baselines in detection probability while achieving up to 43.5× lower computational cost. During idle periods, eventdriven processing further reduces correlation operations by 86% relative to nWUS reception. Index Terms—wake-up radio, neuromorphic computing, lowpower wake-up signal, noncoherent detection, OFDM

I. I NTRODUCTION A wake-up radio (WUR) [1], [2] is an auxiliary low-power receiver that continuously monitors the channel for a specific wake-up signal (WUS) while the main receiver is asleep. WURs can reduce energy consumption by waking up the main receiver only when needed, making them particularly attractive for energy-constrained devices [1]–[3]. To provide meaningful energy savings, both the WUR and the WUS should be designed to minimize the processing required for continuous monitoring and detection [1], [2], [4]. At the same time, orthogonal frequency-division multiplexing (OFDM) is expected to remain a key waveform technology for 6G. Compatibility with standard OFDM transmitters is thus desirable for WUS design [4]–[7]. Neuromorphic computing [8], [9] provides a promising approach for ultra-low-power WURs [10], [11]. Inspired by biological neural networks, neuromorphic systems represent information as discrete binary events called spikes. They process these events in a sparse and event-driven manner [8], i.e., with computation occurring upon spike events. When implemented on neuromorphic hardware [9], this can substantially reduce the computational cost and associated energy

consumption compared with conventional dense processing, especially during idle periods (when the WUR receives only noise). Recent studies have demonstrated the potential of neuromorphic WURs for energy-efficient WUS detection [10]– [12]; for example, a 1.2 nW analog implementation was shown in [10]. However, efficient neuromorphic processing requires inputs represented as discrete spike events [9]. Conventional OFDM waveforms lack this spike-like structure, which limits their direct use with neuromorphic WURs. In this paper, we introduce a neuromorphic wake-up signal (nWUS) that can be generated by a standard OFDM transmitter, as well as a neuromorphic detector that exploits the spikelike structure of the nWUS for efficient WUS detection. The overall method is illustrated in Fig. 1. The main contributions of this paper are as follows. • We propose an OFDM-based nWUS whose time-domain waveform uses spike-like amplitude peaks at configurable positions to encode the wake-up signature, enabling efficient neuromorphic processing at the receiver. • We propose a noncoherent neuromorphic detector that encodes the received amplitude peaks into spikes and performs sparse correlation using a weighted template derived from the target nWUS. The correlator processes only spike events, reducing computation relative to dense correlation and further lowering activity during idle periods when fewer spikes are generated. • We evaluate the nWUS over a 3GPP urban macro channel and compare it with two on-off keying (OOK) baselines of comparable and tenfold duration. Under equal average transmit power, the nWUS improves mean detection probability by 18.6 and 3.3 percentage points (pp) while reducing counted operations by 2.5× and 43.5×, respectively. During idle periods, correlation operations are further reduced by 86% relative to nWUS reception. Overall, this work contributes toward enabling ultra-low-power neuromorphic WURs for OFDM-based wireless systems. II. R ELATED W ORK A. Low-Power Wake-Up Signals Low-power WUSs are designed to enable reliable detection with minimal processing at the WUR [2], [4], [7]. To this end, the WUS should support noncoherent time-domain detection, avoiding carrier synchronization, fast Fourier transform (FFT) operations, and frequency-domain calibration [4], [7],

OFDM transmitter nWUS spike sequence

Subcarrier coefficients

Wireless channel

Transmitted amplitude

Neuromorphic wake-up radio Spike-encoded signal

Received amplitude

Sparse correlator ta tb tc

Weighted template

Λ

t2

t3

F

F

Λ>γ

w[ta] w[tb] w[tc]

IFFT

t1

Decision

t1

t2

t3

F

ta

tb

tc

Lrx

ta

tb

tc

Lrx

t1

t2

t3

Wake up main radio

F

Fig. 1. Overview of the proposed nWUS scheme, comprising OFDM-based nWUS generation and neuromorphic detection. The OFDM transmitter maps the nWUS signature, defined as a spike sequence, to subcarrier coefficients. OFDM modulation then produces sparse amplitude peaks at the spike positions. The receiver encodes the received amplitude peaks into spike events and performs sparse correlation with a template derived from the nWUS signature.

[13]. Existing approaches mainly use OOK, which can be implemented using OFDM transmitters and detected with simple noncoherent envelope-based receivers [3]–[7], [13], [14]. However, in OOK, bits are typically transmitted sequentially, each represented by an ON/OFF symbol occupying a dedicated time interval. This can lead to long WUSs and continuous, dense receiver processing, thereby limiting the energy efficiency achievable at the WUR [4], [6]. Prior work has used energy concentration in OOK to improve detection robustness [14], but combining it with sparsity to lower receiver processing costs has not been investigated.

X ∈ CF into time-domain samples via the F -point inverse fast Fourier transform (IFFT)   F 2π(k − 1)(n − 1) 1 X x[n] ≜ X[k] exp j , (1) F F k=1

where n = 1, . . . , F , giving Ltx = F ; the cyclic prefix is omitted for simplicity. Let y ∈ CLrx denote the received time-domain baseband signal, where Lrx is the observation length. For a timevarying finite-impulse-response channel, the received samples are given by

B. Neuromorphic Wake-Up Radios Neuromorphic computing has recently been explored for WUR applications [10]–[12], [15]–[17]. Existing work mainly considers OOK as the WUS waveform [10]–[12]. OOK can be represented as spike events based on the presence or absence of received energy, but its waveform is neither inherently sparse nor spike-like. In parallel, WURs have been integrated into neuromorphic communication architectures based on impulseradio ultra-wideband (IR-UWB) [16], [17]. IR-UWB signals provide sparse temporal impulses well suited to neuromorphic processing, but are not compatible with conventional OFDM transmitters. To our knowledge, OFDM-based WUS designs that exploit sparse, spike-like time-domain waveforms for neuromorphic detection remain unexplored. III. BACKGROUND Notation: Bold symbols denote column vectors, and a[n] denotes the n-th entry of a. Vector entries and FFT bins are indexed starting at 1, whereas detection lags start at 0. The imaginary unit is denoted by j, with j2 = −1. A. System Model We consider an OFDM-based downlink wake-up scenario. At the transmitter, a WUS is mapped onto an OFDM symbol and transmitted over the air. At the receiver, the WUR searches for a known WUS directly in the time-domain baseband signal. Let x ∈ CLtx denote the transmitted baseband WUS waveform comprising Ltx time-domain samples. For an OFDMbased WUS, a conventional OFDM transmitter with FFT size F converts the frequency-domain subcarrier coefficients

y[n] =

Lh X

h[n, m] x[n − m + 1] + z[n],

(2)

m=1

where Lh is the channel impulse response length, z[n] is additive noise, and h[n, m] is the m-th tap of the discrete-time baseband channel impulse response at sample n; x[n] = 0 outside 1 ≤ n ≤ Ltx . The WUR performs noncoherent rx detection on the received amplitude r ≜ |y| ∈ RL ≥0 . Because the arrival time is unknown, the observation at the WUR is longer than the transmitted waveform: one detection attempt processes Lrx = Ltx + Nlag − 1 samples and evaluates Nlag lags spaced by one sample. Lag ℓ = 0, . . . , Nlag − 1 corresponds to the window of samples ℓ + 1, . . . , ℓ + Ltx . B. Problem Formulation We primarily consider a WUS identification problem in which the receiver must distinguish a target WUS from other valid WUSs. We focus on this setting because it can be more challenging than detection against noise only (both hypotheses contain a valid transmitted signal). Detection is formulated as a binary hypothesis test per attempt: hypothesis H1 denotes reception of the target WUS and H0 denotes reception of a WUS from the same family with a different signature (referred to as a non-target WUS). The WUR computes a score vℓ for each lag window and compares the maximum Λ ≜ max0≤ℓ≤Nlag −1 vℓ with a decision threshold γ ∈ R, selecting H1 if Λ > γ. The detection probability Pd ≜ Pr{Λ > γ | H1 } and the false-identification probability Pfi ≜ Pr{Λ > γ | H0 } are therefore defined per attempt, i.e., over the whole lag search rather than per window, and γ is calibrated accordingly.

In addition to this identification task, we also consider detection against noise only by replacing H0 with Hnoise , under which the received signal in (2) reduces to y[n] = z[n]. The corresponding per-attempt false-alarm probability is Pfa ≜ Pr{Λ > γ | Hnoise }. C. On-Off Keying Baseline We consider an OOK-based WUS, denoted OOK-WUS, whose signature is a binary sequence b̄ ∈ {0, 1}N . A bit “1” is transmitted as an active (ON) interval, whereas a bit “0” is transmitted as an inactive (OFF) interval. Each bit spans Lb ∈ Z>0 time samples, resulting in a waveform of length Ltx = N Lb samples. We generate the waveform directly in the time domain as x[n] = b̄[⌈n/Lb ⌉], without transmit filtering, to obtain an idealized reference. OFDM-based realizations can also be used [13] but introduce additional waveform design choices beyond the scope of this baseline. For noncoherent detection with unknown bit boundaries, the received squared amplitude is integrated over sliding intervals Pp+Lb −1of Lb2 samples, yielding the energy sequence u[p] ≜ r[n] . For each lag ℓ, the corresponding bit-energy n=p vector eℓ ∈ RN ≥0 has entries eℓ [i] ≜ u[ℓ + (i − 1)Lb + 1]. To improve identification performance, eℓ is mean-centered and normalized to unit norm, giving êℓ (if the centered vector has zero norm, êℓ = 0). The receiver then correlates it with the target OOK sequence b̄, producing the per-lag score vℓ ≜ ê⊤ ℓ b̄. Last, the decision follows Section III-B. IV. N EUROMORPHIC WAKE -U P S IGNAL We seek a WUS that is compatible with conventional OFDM transmitters and has a sparse, spike-like time-domain structure, enabling efficient neuromorphic processing at the receiver; we introduce the nWUS to satisfy these requirements. A. Signature Definition The nWUS signature is a length-F spike sequence in which s Ns spike events occur at predefined time indices T ≜ {ti }N i=1 , ordered such that 1 ≤ t1 < · · · < tNs ≤ F . Equivalently, the nWUS signature can be defined as a binary sequence s̄ ∈ {0, 1}F with entries s̄[n] = 1 if n ∈ T and s̄[n] = 0 otherwise. We impose a minimum spacing ∆min ∈ Z>0 between consecutive spikes, such that ti+1 − ti ≥ ∆min . The signature is fixed and known to the WUR during operation; different WURs can be assigned different signatures. Two main design considerations guide the choice of T . First, Ns controls the sparsity of the signature and provides a trade-off between detection robustness and receiver-side computational cost, since more spikes require more operations. Second, ∆min should be large enough to keep the amplitude peaks corresponding to adjacent spikes distinguishable in the generated time-domain waveform. B. Waveform Construction We consider an nWUS occupying one OFDM symbol with FFT size F and D ≤ F active subcarriers. Let B ≜ {κd }D d=1 ⊆ {1, . . . , F } denote the indices of the active

subcarriers, which form a contiguous frequency band. Our goal is to select the subcarrier coefficients so that, after the IFFT, the time-domain waveform exhibits sparse amplitude peaks at the predefined indices in T . For each spike position ti ∈ T , we construct a subcarrier (i) coefficient vector XB ∈ CD with entries   2π(κd − 1)(ti − 1) (i) . (3) XB [d] ≜ exp −j F The linear phase progression makes the active-subcarrier contributions add coherently at sample ti after the IFFT, producing a spike-like amplitude peak. To obtain all spikes in one OFDM symbol, we sum these vectors to form the PNs (i) XB ∈ C D . nWUS subcarrier coefficient vector XB ≜ i=1 Equivalently, XB can be formed by applying an F -point FFT to the spike sequence s̄ and retaining the coefficients at bins κd ∈ B. We emphasize that our contribution lies not in generating time-domain peaks through coherent phase alignment, but in using these peaks and their positions as a sparse WUS signature for efficient neuromorphic processing. We construct the full frequency-domain OFDM symbol X ∈ CF by setting X[κd ] = XB [d] for d = 1, . . . , D and all remaining bins to zero. Here, we choose B to occupy a contiguous band centered around the direct current (DC) subcarrier. An F -point IFFT is then applied to X via (1), producing a time-domain waveform with sparse amplitude peaks at the predefined indices in T . When all F subcarriers are active, the IFFT exactly reconstructs the nWUS spike sequence s̄, with each spike occupying a single time sample. For D < F , the band-limited spectrum produces a Dirichletkernel-shaped pulse around each spike position, with a broader main lobe and deterministic sidelobes. Provided that ∆min is sufficiently large, the peaks should remain distinguishable and their locations should be preserved. The DC subcarrier is kept active since the WUR operates in the time domain, although DC nulling is also supported at the cost of a small deterministic distortion of the time-domain waveform. The entire OFDM symbol is reserved for the nWUS (i.e., not shared with data) to preserve the intended temporal structure. Last, the waveform proceeds through the rest of the OFDM transmission chain. A cyclic prefix is not needed by the nWUS scheme itself, but it can be included when required by the OFDM configuration. In that case, the spike positions should satisfy tNs ≤ F − Lcp , where Lcp denotes the cyclicprefix length, to exclude the peak centers from the samples copied into the prefix. Overall, the nWUS modifies only the frequency-domain coefficients given to the OFDM modulator and thus requires no changes to the OFDM transmitter. C. Information Capacity The nWUS can encode information through the number of spikes and their positions. For an FFT size F , a fixed number of spikes Ns , and the minimum-spacing constraint ti+1 − ti ≥ ∆min , the number of distinct spike sequences is   F − (∆min − 1)(Ns − 1) Nseq ≜ , (4) Ns

where feasibility requires F ≥ 1 + (Ns − 1)∆min . Hence, one nWUS symbol can carry at most CnWUS ≜ ⌊log2 Nseq ⌋ bits. In practice, the usable capacity may be lower due to, for example, ambiguity between shifted signatures under the lag search or limited robustness of some sequences to channel impairments. V. N EUROMORPHIC D ETECTOR We propose a noncoherent neuromorphic detector to efficiently process the nWUS at the receiver. A. Spike Encoding The detector first encodes received amplitude peaks into spike events, converting the continuous signal into a sparse representation that reduces subsequent computational cost. In this work, we use a simple fixed-threshold level-crossing encoder. Encoding is performed once per detection attempt by comparing each of the Lrx received amplitude samples with an encoder threshold η ∈ R>0 . The spike-encoded signal s ∈ {0, 1}Lrx is defined as ( 1, r[n] ≥ η and r[n − 1] < η, s[n] ≜ (5) 0, otherwise, so one spike event is emitted on each rising crossing rather than on every sample above the threshold. The threshold balances detection robustness and computational cost: it must be low enough to detect nWUS peaks but high enough to avoid generating too many spikes during idle periods (when only noise is present). To control this trade-off, we define an average spike budget K ∈ R>0 over the Lrx -sample observation window and select η so that the empirical average spike count under H1 matches this budget. Setting K > Ns allows more spikes than transmitted nWUS peaks, which in practice can improve the robustness of the subsequent detector. Limitations of this spike encoder include its reliance on absolute amplitude levels and the need to calibrate η prior to operation. However, the proposed nWUS and correlator are agnostic to the encoder. Adaptive encoders, such as constantfalse-alarm-rate schemes, could reduce calibration requirements and improve robustness to changing conditions. Also, spike encoding can be performed in the analog domain [10], avoiding high-rate sampling and analog-to-digital conversion. B. Sparse Correlator A sparse correlator matched to the target nWUS signature then processes the spike-encoded signal. The correlator follows two key neuromorphic principles [8]: discrete spike inputs and computation triggered only by spikes. The correlator compares windows of F consecutive spike-encoded samples (matching the transmitted nWUS length) with a template represented by a real-valued weight vector w ∈ RF . Since the nWUS may arrive at an unknown position, the correlation is evaluated at every lag ℓ = 0, . . . , Nlag − 1: vℓ ≜

F X n=1

s[ℓ + n] w[n] =

X 1≤n≤F s[ℓ+n]=1

w[n],

(6)

and the maximum score Λ = max0≤ℓ≤Nlag −1 vℓ is kept. Because only active spike positions contribute to (6), the computational cost of the correlator scales with the number of spikes rather than with the window length. This reduction in computation translates into lower energy usage on hardware that exploits sparsity, such as neuromorphic processors [8], [9]. The weighted template w is constructed directly from the target nWUS signature. All weights are first initialized to a minimum value wmin . Then, for each spike at position ti , the weights are defined for ti ≤ n < ti+1 as   n − ti , (7) w[n] ≜ wmin + (wmax − wmin ) exp − τ where wmax is the maximum value, τ > 0 is the decay constant in samples, and tNs +1 ≜ F + 1. Thus, each spike sets the weight to wmax , which then decays toward wmin . The decay provides tolerance to local temporal distortions within an aligned window, such as band-limiting and multipath smearing. A spike received slightly after its expected position still contributes positively to the score. The decay constant τ controls the trade-off between robustness and temporal selectivity. This tolerance is distinct from the lag search, which absorbs the global delay. Moreover, detection does not require exact recovery of the transmitted spike sequence; the correlation score can remain discriminative when some spikes are missed, temporally shifted, or accompanied by additional unintended spikes. Last, the sparse correlator is hardwareagnostic, supporting implementations in conventional digital circuits, analog circuits, and neuromorphic hardware. VI. S IMULATION R ESULTS We evaluate the proposed nWUS scheme and compare its performance with two OOK-WUS baselines. A. Setup a) nWUS: The nWUS uses one OFDM symbol with an FFT size of F = 1024, D = 870 active subcarriers (85% occupancy, leaving the remaining subcarriers for guard bands), a subcarrier spacing of 15 kHz, and no cyclic prefix (not required here). The resulting sampling rate is 15.36 MHz, and the nWUS duration is 66.67 µs. We set the number of spikes to Ns = 5 and the minimum spike spacing to ∆min = 20 samples; the nWUS signature is generated randomly subject to these sparsity and spacing constraints. The correlator parameters are τ = 3, wmin = −0.1, wmax = 1, and the encoder spike budget is set to K = 14 spikes per observation. The parameters Ns , ∆min , K, τ , and wmin were selected via random search on independent data to balance performance and sparsity. b) OOK-WUS: We use two OOK-WUS baselines from Section III-C to provide complementary comparisons with the nWUS. The first contains N = 1024 bits, matching the length of the nWUS signature. Each bit spans Lb = 10 samples, a duration selected to approach the detection reliability of the nWUS, yielding ten times its airtime (666.67 µs). The second contains N = 42 bits, such that this waveform and the nWUS support approximately the same number of

1.0 0.8 0.6 Pd

distinct WUS signatures (see Section IV-C). Here, Lb = 25 is selected to closely match the duration of the nWUS (OOKWUS duration of 68.36 µs), providing a comparison with similar information capacity and airtime. Both OOK-WUS baselines use the same 15.36 MHz sampling rate as the nWUS. The OOK-WUS signatures are randomly generated with independent equiprobable bits. c) Channel model: We use Sionna 2.0 [18] to simulate the 3GPP TR 38.901 urban macro channel [19]. All waveforms are evaluated over the same single-input single-output downlink at 3.5 GHz. For each realization, user locations, heights ([1.5, 22.5] m), indoor/outdoor conditions (indoor probability of 0.5), and speeds ([0, 30] m/s) are randomly generated. The channel response is normalized to unit energy, and independent complex AWGN is added. The receivers search Nlag = 60 sample-spaced lags to account for channel-induced delays; this gives Lrx = 1,083 for the nWUS, 1,109 for the 42bit OOK-WUS, and 10,299 for the 1024-bit OOK-WUS. No random timing offset is introduced beyond the delay induced by the channel. All waveforms are scaled to the same average transmit power before channel propagation to achieve the target signal-to-noise ratio (SNR).1 d) Evaluation protocol: We evaluate performance through Monte Carlo simulations of the detection problem defined in Section III-B. Target and non-target WUSs use identical parameters and differ only in their signatures, defined by the spike positions for the nWUS and the bit sequence for the OOK-WUS. The target signature is fixed, whereas a non-target signature is drawn independently for every H0 attempt. For the nWUS, we first calibrate a single encoder threshold η by matching the empirical mean spike count to the budget K over 104 independent H1 attempts with random SNRs in the evaluated range. After this preliminary step, the nWUS and the OOK-WUS baselines are evaluated using the same procedure. At each SNR point, we calibrate the decision threshold γ from 105 independent H0 attempts by selecting the empirical quantile corresponding to the target false-identification probability Pfi = 10−2 . Using the resulting threshold, we estimate the detection probability Pd from 105 independent H1 attempts and verify the achieved Pfi on a separate set of 105 independent H0 attempts. For the secondary noise-only problem, we replace H0 with Hnoise throughout this procedure: γ is calibrated using noise-only attempts for a target Pfa = 10−2 , while all other settings remain unchanged. To assess sensitivity to the WUS signature, we repeat the experiment for five random seeds, each defining a different target signature. The reported results show the mean across these five runs.

0.4 nWUS (F = 1024, Ns = 5) OOK–WUS (N = 1024, Lb = 10)

0.2 0.0

OOK–WUS (N = 42, Lb = 25) Target (Pd = 0.99)

-14 -12 -10

-8

-6

-4 -2 0 SNR (dB)

2

4

6

8

10

Fig. 2. Detection probability Pd versus SNR under equal average transmit power, with the decision threshold calibrated to achieve Pfi = 10−2 .

Pd = 0.99 as the target detection reliability. The achieved Pfi remains in the range [0.93, 1.11] × 10−2 . Compared with the 42-bit OOK-WUS, which has a similar waveform duration and information capacity, the nWUS achieves a substantially higher Pd over the entire SNR range, with a gain of 18.6 pp averaged over all evaluated SNR points. The 1024-bit OOK-WUS is ten times longer than the nWUS and therefore transmits ten times more energy at the same average transmit power. Despite this advantage, it is outperformed by the nWUS, which achieves a mean Pd gain of 3.3 pp. The nWUS reaches the target Pd at an SNR of approximately −8 dB, compared with approximately 0 dB for the 1024-bit OOK-WUS, while the 42-bit OOK-WUS remains below the target throughout the evaluated SNR range. Overall, these results show that the proposed nWUS achieves higher detection performance than OOK-WUS with comparable or lower airtime. We attribute the performance gain primarily to two factors: (i) concentrating the available WUS energy into sparse temporal peaks, which improves spike recovery at the receiver, and (ii) using a weighted correlator template tailored to spike sequences, which also improves robustness to local, channel-induced spike shifts. Finally, we evaluate detection against noise only. In this setting, the nWUS reaches Pd = 0.99 at −10 dB, compared with −6 dB for the 1024-bit OOK-WUS and 0 dB for the 42bit OOK-WUS. These results show that the nWUS retains its detection advantage over both baselines when distinguishing the target WUS from noise only.

B. Detection Performance

C. Computational Cost

Fig. 2 shows the detection probability as a function of SNR for the proposed nWUS and the two OOK-WUS baselines, all evaluated under equal average transmit power. We use

A main motivation for the nWUS is to reduce the computational cost at the receiver through sparse, event-driven computing. We assess this by approximating the number of operations required for one detection attempt. For the nWUS detector, the level-crossing encoder performs one comparison per sample, giving Lrx preprocessing operations. The sparse

1 We define SNR as the ratio of the average transmitted signal power per sample (before the channel) to the noise power per complex sample.

TABLE I E STIMATED COMPUTATIONAL COST PER DETECTION ATTEMPT. Operations Scheme

Preproc.

Corr.

Total

Rel. cost

nWUS (Hnoise ) nWUS (H1 ) 42-bit OOK-WUS (H1 ) 1024-bit OOK-WUS (H1 )

1,083 114 1,197 1,083 804 1,887 2,218 2,520 4,738 20,598 61,440 82,038

0.6× 1× 2.5× 43.5×

correlator performs one addition per spike in each lag window. Its total cost per detection attempt is therefore Nlag times the average number of spikes per window, measured as 13.4 under H1 and 1.9 under Hnoise . For the OOK-WUS detector, the running sum providing the bit energies requires 2Lrx preprocessing operations (one addition and one subtraction per sample). The dense correlator requires N × Nlag additions; it is kept dense (without zero skipping) to isolate the benefits of sparse computation on appropriate hardware. We omit the mean-centering and unit-norm normalization used to improve performance because their computational cost is strongly implementation-dependent and they have no direct counterpart in the nWUS detector. Each comparison, addition, or subtraction is counted as one operation, as their energy costs are comparable [20]. Table I reports the number of counted operations for each scheme. At comparable airtime and information capacity, the 42-bit OOK-WUS requires 2.5× as many counted operations as the nWUS under H1 . This ratio increases to 43.5× for the 1024-bit OOK-WUS, whose tenfold airtime further increases the processing cost. During noise-only periods (Hnoise ), the nWUS correlator performs approximately 86% fewer operations than under H1 , while the OOK-WUS correlator incurs the same cost regardless of received activity. This illustrates the computational benefits of sparse, event-driven processing. Together with the findings in Section VI-B, these results show that the nWUS combines computational savings with higher detection probability relative to OOK-WUS baselines of both comparable and substantially longer duration. We note that the reported operation counts should be interpreted as coarse proxies for relative computational cost and associated receiver energy, as they omit additional operations and implementationdependent costs such as memory access. VII. C ONCLUSION At a high level, both an nWUS and an OOK-WUS encode a bit-like temporal signature that can be detected using correlation. Their main difference lies in how this signature is represented in the time-domain waveform. In an OOK-WUS, active symbols are frequent and extend over multiple time samples, whereas in an nWUS, active symbols occur sparsely and are localized to one or a few time samples. This sparse representation enables a temporally compact waveform that concentrates transmit energy into fewer active samples. At the receiver, energy concentration improves detectability, sparsity reduces the computational cost of correlation, and event-driven

processing further reduces this cost during idle periods. These benefits are obtained while retaining OFDM compatibility, making the proposed nWUS a promising waveform for ultralow-power neuromorphic WURs in OFDM-based wireless systems. More broadly, the nWUS is not tied to the proposed detector and could also benefit existing neuromorphic WUR architectures that rely on spike-based WUS signatures. Realizing these benefits in practical systems requires further investigation of hardware constraints and integration aspects. In particular, the impact of the high amplitude peaks on power amplifier operation remains unassessed and should be studied in future work. We will also investigate multiplexing the nWUS with user data on unused subcarriers to improve spectral efficiency. R EFERENCES [1] R. Piyare, A. L. Murphy, C. Kiraly, P. Tosato, and D. Brunelli, “Ultra low power wake-up radios: A hardware and networking survey,” IEEE Commun. Surveys Tuts., vol. 19, no. 4, 2017. [2] P. P. Mercier et al., “Low-power RF wake-up receivers: Analysis, tradeoffs, and design,” IEEE Open J. Solid-State Circuits Soc., vol. 2, 2022. [3] D.-J. Deng et al., “IEEE 802.11ba: Low-power wake-up radio for green IoT,” IEEE Commun. Mag., vol. 57, no. 7, 2019. [4] 3GPP, “Study on low-power wake-up signal and receiver for NR,” TR 38.869 V18.0.0, Release 18, 2024. [5] N. Mazloum and O. Edfors, “Interference-free OFDM embedding of wake-up signals for low-power wake-up receivers,” IEEE Trans. Green Commun. Netw., vol. 4, no. 3, 2020. [6] A. Höglund, M. Mozaffari, Y. Yang, G. Moschetti, K. Kittichokechai, and R. Nory, “3GPP Release 18 wake-up receiver: Feature overview and evaluations,” IEEE Commun. Standards Mag., vol. 8, no. 3, 2024. [7] S. Wagner, “Low-power wake-up signal design in 3GPP 5G-Advanced Release 19,” 2026, arXiv:2507.10207v2. [8] C. D. Schuman et al., “A survey of neuromorphic computing and neural networks in hardware,” 2017, arXiv:1705.06963. [9] M. Davies et al., “Advancing neuromorphic computing with Loihi: A survey of results and outlook,” Proc. IEEE, vol. 109, no. 5, 2021. [10] Z. Jouni, T. Soupizet, S. Wang, A. Benlarbi-Delai, and P. M. Ferreira, “1.2 nW neuromorphic enhanced wake-up radio,” in Proc. SBC/SBMicro/IEEE/ACM Symp. Integr. Circuits Syst. Design, 2022. [11] G. Marthe, C. Goursaud, and L. Clavier, “Wake-up radio receiver based on spiking neurons for detecting activation sequence,” in Proc. IEEE Wireless Commun. Netw. Conf., 2023. [12] Z. Jouni, T. Soupizet, S. Wang, A. Benlarbi-Delai, and P. M. Ferreira, “RF neuromorphic spiking sensor for smart IoT devices,” Analog Integr. Circuits Signal Process., vol. 117, 2023. [13] D.-J. Deng, S.-Y. Lien, C.-C. Lin, M. Gan, and H.-C. Chen, “IEEE 802.11ba wake-up radio: Performance evaluation and practical designs,” IEEE Access, vol. 8, 2020. [14] R.-A. Pitaval and X. Tie, “DFT-s-OFDM-based on–off keying for lowpower wake-up signal,” IEEE Trans. Commun., vol. 73, no. 10, 2025. [15] G. Marthe, C. Goursaud, and L. Clavier, “Using synapse saturation in spiking neural networks for wake up receivers in Internet of Things networks,” IEEE Internet Things J., vol. 12, no. 21, 2025. [16] J. Chen, S. Park, P. Popovski, H. V. Poor, and O. Simeone, “Neuromorphic split computing with wake-up radios: Architecture and design via digital twinning,” IEEE Trans. Signal Process., vol. 72, 2024. [17] J. Chen, S. Park, P. Popovski, H. V. Poor, and O. Simeone, “Neuromorphic semantic communications with wake-up radios,” in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun., 2024. [18] J. Hoydis et al., “Sionna: An open-source library for next-generation physical layer research,” 2022, arXiv:2203.11854. [19] 3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,” TR 38.901 V19.4.0, Release 19, 2026. [20] F. Manzouri et al., “A comparison of energy-efficient seizure detectors for implantable neurostimulation devices,” Front. Neurol., vol. 12, 2022.

Related documents

• OFDM-Based Neuromorphic Wake-Up Signal#1468518
• Spiking neural networks provide accurate and time-efficient models for whisker stimulus classification of the awake mouse.#15181
• FEDERATED NEUROMORPHIC INTELLIGENCE: ADVANCING ROBUSTNESS, EFFICIENCY, AND CONTINUAL ADAPTATION IN EDGE ENVIRONMENTS#226453
• Edge AI on Constrained Devices for Binary Sleep-Wake Classification in Dynamic Environments#1077092
• End-to-End Radar and Communication Modulation Recognition with Neuromorphic Computing#1399901